Digital Soil Texture Mapping and Spatial Transferability of Machine Learning Models Using Sentinel-1, Sentinel-2, and Terrain-Derived Covariates

Abstract

Soil texture is an important property that controls the mobility of the water and nutrients in soil. This study examined the capability of machine learning (ML) models in estimating soil texture fractions using different combinations of remotely sensed data from Sentinel-1 (S1), Sentinel-2 (S2), and terrain-derived covariates (TDC) across two contrasting agroecological regions in Southwest Germany, Kraichgau and the Swabian Alb. Importantly, we tested the predictive power of three different ML models: the random forest (RF), the support vector machine (SVM), and extreme gradient boosting (XGB) coupled with the remote sensing data covariates. As expected, ML model performance was not consistent regarding the input covariates, soil texture fractions, and study regions. For example, in the Swabian Alb, the SVM model performed the best for the sand content with S2 + TDC (RMSE = 3.63%, R2 = 0.42), and XGB best predicted the clay content with S1 + S2 + TDC (RMSE = 6.84%, R2 = 0.64). In Kraichgau, the best models for sand (RMSE = 7.54%, R2 = 0.79) and clay contents (RMSE = 6.14%, R2 = 0.48) were obtained using XGB and SVM, respectively. Moreover, the results indicated that TDC were critical in estimating soil texture fractions, especially in Kraichgau, which indicated that topography plays an important role in defining the spatial distribution of soil properties. In contrast, the contribution of remote sensing data better predicted the silt and clay content in the Swabian Alb. The transferability of a region-specific model to the other region was low as indicated by poor predictive performance. The resulting soil-texture-fraction maps could be a significant source of information for efficient land resource management and environmental monitoring. Nonetheless, further research to evaluate the added value of the Sentinel imagery and to better analyze the spatial transferability of machine learning models is highly recommended.

An assessment of Sentinel‐1 synthetic aperture radar, geophysical and topographical covariates for estimating topsoil particle‐size fractions

“…As a first soil layer, it is more prone to undergo surface interactions than subsurface interactions. In relation to the topography covariates being identified as important predictors and machine learning models (e.g., RF) not performing well in some situations, these general findings are consistent with Mirzaeitalarposhti et al (2022) and Schönbrodt‐Stitt et al (2021) who also applied Sentinel‐1 data and topography covariates to estimate soil particle‐size fractions and soil moisture.…”

“…Random forest (RF) (Breiman, 2001) is one example of treebased models, and it aims to merge less powerful learners to form a strong learner to minimize the residual sum of squares by tuning two main hyper parameters-the number of trees (ntree) and the number of features randomly sampled at each split (mtry). Successful examples of RF applications to estimate soil properties-without considering soil texture as compositional data-in remote sensing context are found in Mirzaeitalarposhti et al (2022), Domenech et al (2020), Dotto et al (2020, Cisty et al (2019), Bousbih et al (2019) and Ballabio et al (2016).…”

Application of Machine Learning Algorithms for Digital Mapping of Soil Salinity Levels and Assessing Their Spatial Transferability in Arid Regions

“…Therefore, it is recommended for future studies to present comparative results of different mathematical bases such as the Gower similarity index [93] and dissimilarity index [94]. Enhancing the predictive accuracy of transferability related to soil salinity can involve the exploration of specific geographical stratifications [53], such as physiography or topography (slope-aspect categories), as well as the consideration of land use factors.…”

“…The transfer learning approach has been demonstrated to be applicable in pedometrics, especially in studies on the prediction of soil properties by creating and using spectral reflectance libraries [45][46][47]. By integrating the transfer learning approach, relevant DSM studies were conducted, such as the parent material [48], organic carbon at the local scale [49], USDA Soil Taxonomy at the sub-group level [50], USDA Soil Taxonomy at the soil great group level [51], soil organic carbon in cropland soils [52], and soil particle fractions [53].…”

Identification of Soil Types and Salinity Using MODIS Terra Data and Machine Learning Techniques in Multiple Regions of Pakistan

“…As a result, it becomes essential to assess the performance of different models when dealing with multisource datasets, especially in domains characterized by significant intrinsic and extrinsic variability [25]. Soil type and crop yield estimation may be possible with better accuracy at higher geographic resolutions using a method that simultaneously combines multisensor RS data along with environmental variables [30][31][32].…”